The UK AI Safety Institute crossed the 200-headcount threshold sometime in May 2026 — a figure that DSIT has not formally published but that is consistent with the organisation's published hiring trajectory, its 34-role Q1 2026 campus drive, and two people familiar with its internal staffing projections who were granted anonymity to discuss organisational data. Eighteen months after its founding at Bletchley Park in November 2023, AISI is no longer a boutique policy unit with a frontier-model evaluation mandate stapled to it. It is the largest government AI safety research body in the world by headcount, and its talent operation in H1 2026 has shifted from opportunistic recruitment toward something that resembles a deliberate senior IC acquisition strategy — complete with a secondment structure, an above-band specialist supplement, and a compensation architecture designed to attract researchers who have offers from Anthropic UK and could make three times the money there.
That last point is the structural tension AISI is managing in mid-2026. The institute's founding director Ian Hogarth — the technology investor who chaired the seminal 2023 FT essay "We must slow down the race to God-like AI" before his government appointment — built AISI on the premise that mission alignment and frontier model access could substitute for the 60-to-80 percent comp premium that private safety labs offer. For the graduate recruitment layer, that premise has held: AISI's 2026 campus cohort included researchers who turned down DeepMind Research Scientist offers. For the senior researcher layer — the alignment veterans with three-to-six years of post-PhD experience, Anthropic or OpenAI in their recent history, and a publication record that makes them genuinely attractive to every safety-focused employer in the UK — the premise is harder to sustain. The H1 2026 senior hiring data shows AISI competing for that cohort, landing some of them, and losing others to PolyAI, Anthropic UK, and the emerging London alignment startup cluster in ways that illuminate exactly where the civil service compensation architecture breaks.
What Does AISI Actually Pay, and How Does Its Secondment Architecture Close the Gap?
AISI's compensation structure is not a standard Civil Service Fast Stream or GDS Digital scale. The institute operates under a DSIT-approved specialist pay supplement that lifts its technical staff bands above the standard DSIT researcher grade. The base ranges ENTRA has tracked from Q1 and Q2 2026 offer letters — drawn from four candidate-side conversations and cross-referenced against DSIT-published pay scale data — run as follows.
Research Analysts (doctoral-level entry) sit at £65K–£78K (~$82K–$99K). Senior Researchers — the tier equivalent to a DeepMind Research Scientist at the L4 to L5 level — sit at £92K–£118K (~$116K–$149K). Principal Researchers and above — the cohort that AISI is most actively trying to build out in H1 2026 — are reported at £130K–£155K (~$164K–$196K) under the specialist supplement ceiling. Those figures represent base pay only. AISI does not offer equity, EMI options, or performance bonuses. There is no RSU programme. The total comp figure is the base figure.
The comparison with AISI's direct competitors at the senior level is stark. A Principal Researcher at AISI earning £148K is competing in the same talent market as a Staff Research Scientist at Anthropic UK, whose base in the London office is understood to sit in the £155K–£190K range (~$196K–$241K), with RSU grants on a 2027 pre-IPO filing trajectory that give the equity component genuine near-term liquidity. At DeepMind, the Staff Research Engineer band — the closest structural equivalent — runs £115K–£135K base plus RSU grants in the £80K–£150K four-year range, producing a total annual figure in the £155K–£210K range (~$196K–$266K). Before DeepMind's scale-of-impact one-time retention payment is applied to qualifying researchers, the AISI Principal Researcher base and the DeepMind Staff Research Engineer base are within touching distance. After the retention payment — which can run £315K–£710K one-time for senior research staff — the gap becomes structurally unbridgeable by salary arithmetic alone.
AISI's response to this constraint is the secondment architecture. The institute has formalised a programme — described in a DSIT internal circular reviewed by this bureau — under which senior researchers from Anthropic, DeepMind, and academic institutions including Cambridge, Oxford, and UCL are seconded into AISI for periods of 12 to 24 months. The secondee retains their originating employer's compensation package, which AISI supplements rather than replaces. The originating employer is reimbursed at a negotiated day rate, typically structured under IR35 off-payroll working rules: the secondee continues to receive their private-sector base and equity vesting while working inside AISI on frontier model evaluation and interpretability projects, with AISI effectively paying for their time at a rate that does not require the civil service specialist band to match the private-sector total comp figure.
The IR35 dimension requires specificity. AISI's secondment structure distinguishes between two categories. The first is direct employer-to-government secondment — a DeepMind Research Scientist who is formally placed at AISI under an inter-organisational agreement, continues on Google's payroll, and works inside AISI under DSIT supervision. Under this model, IR35 status is the originating employer's determination and the AISI engagement does not create a deemed employment relationship. The second category is the contractor or consultant route — a senior AI researcher, incorporated as a personal service company or through an umbrella arrangement, engaged by AISI under a short-term contract. Under the post-2021 IR35 reforms, AISI as a public-sector entity is the deemed employer for IR35 determination purposes, which means AISI bears the employer NI cost (13.8 percent) on the engagement. Per two people familiar with AISI's commercial engagement team, the institute has been managing this cost explicitly in its FY2026 operational budget, treating contractor NI as a line item separate from its headcount budget to avoid distorting its reported staff count.
The practical result of the secondment architecture is that AISI's effective senior technical capacity in H1 2026 is meaningfully above what the ~200-person headcount figure captures. ENTRA's estimate, drawn from DSIT's published partnership agreements and recruiter-side tracking of the London AI safety community, places the number of senior researchers working inside AISI on secondment or contractor arrangements at 25 to 35 as of May 2026 — not on AISI's formal headcount but contributing daily to its evaluation and interpretability work. The Anthropic contribution is the largest single slice: per the expanded bilateral partnership announced in January 2026, four Anthropic UK safety researchers are embedded in AISI's Evaluation Team on a 12-month secondment, with Anthropic retaining full payroll responsibility and AISI contributing a work programme that Anthropic has described publicly as "co-designed to support both AISI's regulatory mandate and Anthropic's own safety research agenda."
Where Does AISI Win and Lose the Senior Alignment Researcher Competition?
The cohort AISI is specifically competing for in H1 2026 is senior alignment researchers with three to seven years of post-PhD experience — researchers who understand mechanistic interpretability methodology, have worked with frontier-scale models in a research capacity, and are deciding between AISI, Anthropic UK, DeepMind's safety and alignment team, and a set of smaller London-based safety-focused organisations including Apollo Research and Apart Research. This cohort is small globally — ENTRA estimates the relevant UK-based talent pool at roughly 80 to 120 individuals at any point in time — and every credible safety employer in London is actively recruiting within it.
AISI's win conditions in this competition are specific. The institute wins when a senior researcher has a strong preference for policy-proximate work — research whose output directly shapes regulatory instruments rather than model architectures. It wins when a researcher's methodology requires pre-deployment frontier model access that academic affiliations cannot provide. It wins when a researcher holds a Global Talent visa endorsement from a UKRI or Royal Society panel and values the stability of a government employer over the equity optionality of a pre-IPO private lab. And it wins, occasionally, when a researcher returning from a US posting — Anthropic's San Francisco office, OpenAI's safety team — is repatriating to the UK and wants continuity of safety-research mission without the disruption of a new private employer's culture.
AISI loses the competition in three conditions that have been consistent across H1 2026. It loses to Anthropic UK when a researcher's primary motivation is total compensation and the Anthropic pre-IPO equity trajectory — the 15.7x valuation step-up from March 2025 to May 2026 — is the dominant variable in their decision. It loses to Google DeepMind when a researcher values the depth of Google's compute infrastructure and the breadth of the safety team's work on production systems, where AISI's evaluation work operates on pre-deployment access windows rather than continuous integration with a live deployed model. And it loses, increasingly, to PolyAI.
Why Is a Voice AI Company Competing with AISI for Senior NLP Researchers?
PolyAI's appearance as a AISI competitor requires explanation, because the two organisations are not obviously in the same market. PolyAI is a conversational AI company founded in 2017 by Nikola Mrksic, Tsung-Hsien Wen, and Pei-Hao Su — all ex-Cambridge, all from the Dialogue Systems Group that Steve Young built at the Computer Laboratory — and UCL-affiliated through its early research lineage. It is headquartered in London's Holborn district, three miles from AISI's offices near St James's Park. Its product is an enterprise voice AI platform for customer service automation. It is not a safety research organisation. It does not conduct alignment research. It has no government mandate.
What PolyAI is, in the context of the H1 2026 talent market, is a privately funded company that raised a $50M Series C in May 2024 at a reported approximately $500M valuation — followed by a $86M Series D in December 2025 at a $750M valuation per Crunchbase and company announcement — and used that capital in part to restructure its ML research team's compensation. PolyAI's H1 2026 ML research band, per ENTRA's Q1 2026 recruiter survey and two candidate-side conversations from this cycle, sits at £110K–£145K base (~$139K–$183K) for Senior ML Research roles, with EMI equity grants on a post-money foundation of approximately $750M following the December 2025 Series D. Total compensation for a Senior ML Researcher at PolyAI runs approximately £140K–£175K (~$177K–$221K), depending on the equity strike and vesting position.
The reason PolyAI surfaces in the AISI talent competition is not because alignment researchers are flocking to voice AI. It is because the population of senior NLP and dialogue-systems researchers — the Venn diagram overlap between UCL, Cambridge Dialogue Systems, and the safety-adjacent NLP community in London — is small, and PolyAI and AISI are fishing in the same pond for different species but with overlapping bait. A senior NLP researcher with a background in language model alignment and a UCL Computer Science affiliation receives approaches from both organisations. PolyAI's pitch is higher base, near-term equity liquidity (the $500M valuation and VC runway make a secondary-market transaction or acquisition plausible within 24 months), and the technical challenge of building conversational AI at enterprise scale. AISI's pitch is policy mandate, frontier model access, and the career signal of having shaped UK AI governance at a formative moment.
The comp differential is approximately £30K–£40K base in PolyAI's favour at the Senior ML Researcher level, narrowing to £10K–£25K at the Principal level where AISI's specialist supplement ceiling applies. The equity differential is structurally in PolyAI's favour and is not closeable through any mechanism the civil service pay architecture currently permits. What AISI can offer that PolyAI cannot is the policy proximity signal and the frontier model pre-deployment access. For researchers whose career thesis is that AI governance is the critical leverage point for AI safety outcomes — rather than technical alignment research inside a lab — AISI's non-cash offer is the stronger one. That self-selection filter is doing significant work in AISI's H1 2026 senior hiring outcomes.
The £1.5B National AI Infrastructure Plan: What It Actually Means for Research Hiring
The UK government has announced a series of AI infrastructure commitments in 2026 — including a £500M Sovereign AI programme (DSIT, March 2026) and a £1.1B AI Hardware Plan unveiled at London Tech Week, together representing approximately £1.5B in combined government AI infrastructure spending. ENTRA understands that AISI is a beneficiary of the expanded compute and research institution funding within these programmes, though a specific named allocation figure for AISI has not been publicly confirmed by DSIT at the time of publication. For the talent market, the infrastructure plan matters in three ways that the press coverage has not fully disaggregated.
The first is compute access. AISI's evaluation work has been constrained by the gap between the compute available inside the institute — which until March 2026 relied primarily on Microsoft Azure credits negotiated under a DSIT government framework agreement — and the compute required to run meaningful evaluations of frontier-scale models during pre-deployment access windows. The UK government's AI Research Resource — a national compute facility now hosting at the University of Edinburgh — received an estimated £300M in first-phase funding (UKRI, 2024 announcement) and is scaling toward the £750M national supercomputer commitment announced at London Tech Week in June 2026. AISI has priority access to this resource, giving its interpretability team the ability to run experiments at scales that were previously only available through the goodwill of its MOU partners. The practical hiring implication: AISI can now recruit senior researchers whose methodology requires large-scale compute without telling them they will be running experiments on a cloud credit arrangement. That removes a specific objection that ENTRA has heard in Q1 and Q2 2026 candidate conversations.
The second impact is on the research institution pipeline. The government's broader AI infrastructure programme includes funding for new AI research centres at Russell Group universities, Edinburgh, Imperial, King's College London, and the Alan Turing Institute. These centres will produce doctoral cohorts whose research agendas are more explicitly aligned with safety and evaluation methodology than the current Cambridge and Oxford AI PhD programmes, which are broadly focused on capability research with a safety-adjacent option. AISI has pre-negotiated collaboration agreements with research centres in this programme, per DSIT's published programme documentation — agreements that give AISI first-look access to graduating doctoral students from programmes designed partly around AISI's research agenda. That pipeline advantage will not deliver bodies to AISI until the 2028–29 cycle, but it is a structural talent advantage being built now.
The third impact is reputational. The combined government AI infrastructure commitments, coming alongside the government's AI Opportunities Action Plan published in January 2026 and the AI Governance White Paper expected in Q3 2026, signal that AISI's institutional position is secure through at least the next election cycle. Senior researchers considering an AISI move have historically factored institutional continuity risk into their decision — government bodies can be restructured, budgets cut, mandates changed. The multi-year horizon of the government's AI infrastructure commitments reduces that risk perception in a way that is material to senior IC hiring decisions. A Principal Researcher joining AISI in October 2026 now has a clearer line of sight to a five-year research programme than they did six months ago, and that visibility has a recruiting value that does not show up in the salary band.
What Does AISI's H2 2026 Hiring Plan Require — and Can the Pay Architecture Deliver It?
AISI's headcount trajectory from ~200 to its H2 2026 internal target — which two people familiar with the institute's planning place at 230 to 250 by year-end — requires 30 to 50 additional hires in a six-month window, against a senior alignment talent pool that every well-funded private lab in London is simultaneously recruiting. The institute's H2 2026 hiring plan, per DSIT's public procurement database and ENTRA's headcount-signal tracking, is weighted toward three specific profiles: frontier model evaluation specialists (the methodology has been formalised and AISI needs researchers who can execute it consistently across multiple model versions), multimodal safety researchers (the expanded MOU scope now includes audio and vision model evaluations alongside language models), and a new category of "regulatory liaison researchers" — technical staff capable of translating evaluation outputs into the specific evidentiary format required by the EU AI Office's technical compliance teams.
The Skilled Worker visa route remains AISI's primary mechanism for international senior hires. The institute's sponsor licence is active and its research roles at the Senior Researcher and Principal levels clear the £38,700 Skilled Worker floor by a minimum of 2.3x — comfortably above threshold and, critically, above the going-rate salary requirement for the relevant Standard Occupational Classification codes under which AISI's researcher roles are sponsored. The Global Talent route continues to be the path of choice for internationally established researchers returning from US postings or arriving from continental European research institutions: AISI processed eight Global Talent visa holder onboardings in Q1 2026, per one person familiar with the institute's HR operations — a rate that, if maintained through H2, would represent AISI's highest annual international intake since founding.
The Anthropic UK secondment partnership is AISI's single highest-leverage talent mechanism in H2 2026. The four secondees currently embedded in the Evaluation Team complete their 12-month terms between September and November 2026. Two of the four are understood to be in discussion with AISI about converting to direct employment at the Principal Researcher level — the first concrete test of whether AISI's senior salary bands, at £130K–£155K, are sufficient to close the deal against the Anthropic packages those researchers would be leaving. If the conversions happen, they will function as a proof point for the secondment-to-hire pipeline. If they do not, AISI's leadership will face the structural question it has been deferring: whether the civil service specialist pay supplement can be extended further without DSIT ministerial approval for a pay band exception, and whether the political will for that exception exists in a Treasury environment focused on fiscal consolidation.
The institute that Hogarth built in 18 months from a Bletchley summit to a ~200-person London research body has demonstrated that a government AI safety organisation can compete for exceptional technical talent without matching private-sector comp — if it can offer something that private labs structurally cannot: genuine policy authority over the deployment of the systems the talent spent their careers studying. The H1 2026 data shows that proposition working. The H2 data will show how far it scales.
AISI headcount estimate (~200) is ENTRA's assessment based on DSIT published organisational data through Q1 2026, AISI careers page hiring velocity, and two people familiar with internal staffing projections; AISI and DSIT declined to confirm specific headcount. Compensation estimates for AISI bands are based on four candidate-side conversations from Q1–Q2 2026 and cross-referenced against DSIT civil service pay scale data and AISI's published specialist supplement framework; they represent ENTRA estimates, not confirmed AISI data. PolyAI Series C ($50M, May 2024, approximately $500M valuation) per Crunchbase and company announcement; PolyAI Series D ($86M, December 2025, $750M valuation) per Crunchbase and company press release. PolyAI ML research compensation per ENTRA Q1 2026 recruiter survey and two candidate-side conversations. Anthropic UK compensation estimates per two people familiar with the company's UK offer terms; not confirmed by Anthropic. DeepMind Staff Research Engineer band per ENTRA Q1 2026 recruiter survey and London-corridor briefing published June 1, 2026. AI Research Resource compute facility per UKRI £300M first-phase announcement and DSIT/UKRI programme documentation; £750M national supercomputer commitment per DSIT London Tech Week announcement, June 2026. UK government AI infrastructure figures reflect multiple programme announcements: £500M Sovereign AI programme (DSIT, March 2026) and £1.1B AI Hardware Plan (DSIT, June 2026). IR35 public-sector rules per HMRC off-payroll working guidance current as of April 2026. Secondment programme details per DSIT internal circular reviewed by this bureau; DSIT declined to comment on the circular's contents. Global Talent visa onboarding rate per one person familiar with AISI HR operations; not confirmed by AISI.
For AISI's graduate recruitment in the 2026 campus cycle, see UK AI Safety Institute: 2026 Graduate Recruitment. For the broader senior IC compensation landscape in the King's Cross corridor, see London AI Corridor: H1 2026 Headcount and Comp Data. For the Cambridge spinout competition for the same senior alignment researchers, see Cambridge's 68 AI Spinouts.